# FLAS: Flow-based Activation Steering for Inference-Time Intervention

[![Project Page](https://img.shields.io/badge/Project%20Page-flas--ai.github.io-4c1)](https://flas-ai.github.io)
[![arXiv](https://img.shields.io/badge/arXiv-2605.05892-b31b1b.svg)](https://arxiv.org/abs/2605.05892)
[![🤗 Model](https://img.shields.io/badge/🤗%20Model-flas--gemma--2-FFD21E)](https://huggingface.co/collections/flas-ai/flas)
[![🤗 Demo](https://img.shields.io/badge/🤗%20Spaces-flas--demo-FFD21E)](https://huggingface.co/spaces/Lunamos/flas-demo)

Official organization for FLAS.

FLAS is the Flow-based Activation Steering which learns a general, concept-conditioned velocity field that transports unsteered activations to steered ones for the inference-time intervention of LLMs.

## Links

*   **Code Repository:** [FLAS](https://github.com/flas-ai/FLAS)
*   **Project Homepage:** [https://flas-ai.github.io](https://flas-ai.github.io)
*   **Paper:** [Read more on arXiv](https://arxiv.org/abs/2605.05892)
*   **Demo:** Try FLAS on [Hugging Face Space](https://huggingface.co/spaces/Lunamos/flas-demo)